Indeed. LLMs like Chat GPT are pretty hopeless at logic and arithmetic since they can only regurgitate stuff that they have seen in their training. Their natural language abilities are pretty amazing though.
You can work out the miss distance of 3I assuming no gravitational accelerations by linearly extrapolating the “infinity” state vector to closest approach. JPL Horizons will give you the barycentric position and velocity vectors at any past time if you select “vector table” as the ephemeris type. The state vector for 1900-01-01 is:
2415020.500000000 = A.D. 1900-Jan-01 00:00:00.0000 TDB
X = 6.146692847218195E+02 Y =-1.410989916978468E+03 Z = 6.302147691517926E+01
VX=-1.339387085953456E-02 VY= 3.065893078363736E-02 VZ=-1.367393155520660E-03
(position in au, velocity in au/day). The position at any future time (if you ignore gravity) is just P + Vt where P and V are the initial position and velocity vectors. Differentiate the magnitude of that and find the minimum to get the closest approach time (JD 2461021) and you get a miss distance of 1.60au so you did eventually get the correct answer! You should ask it how it knew this? It will have found it on a website somewhere. I’ll be more impressed when these “AI” sites can actually calculate stuff from first principles.